Forecasting Forward Patent Citations: Comparison of Citation-Lag Distribution, Tobit Regression, and Deep Learning Approaches

This article aims to compare the existing approaches, such as citation-lag distribution, Tobit regression, and multilayered neural network to forecast forward patent citations as a proxy for patent technological impact. For the purpose of analysis, in this article, we focuse on estimating lifespan forward citations in the telecommunications sector because sector-specific characteristics may affect the citation patterns. Accordingly, U.S. utility patents in the telecommunications sector are collected and the patent ex-ante information (the information provided at the point of patent disclosure) along with patent ex-post information (the information available after patent disclosure) are extracted to be used for forecasting forward citations. The research findings indicate that the deep learning approach tends to outperform the other approaches only with the ex-ante information, and the performance of the deep learning approach increases significantly when the ex-post information is added to the ex-ante information. This article is expected to offer useful guidelines for choosing an appropriate approach to forecast citations based on available dataset.